Rule-Based Automated Price Negotiation: an Overview

and an Experiment

Costin Bădică1 , Maria Ganzha2 , and Marcin Paprzycki3

1
University of Craiova, Software Engineering Department
Bvd.Decebal 107, Craiova, RO-200440, Romania
badica_costin@software.ucv.ro
2
Elbla˛g University of Humanities and Economy, ul Lotnicza 2, 82-300 Elbla˛g, Poland
ganzha@euh-e.edu.pl
3
Computer Science Institute, Warsaw School of Social Psychology, 03-815 Warsaw, Poland
Marcin.Paprzycki@swps.edu.pl

Abstract. Recently, the idea of automating e-commerce transactions attracted a

lot of interest. Multi-agent systems are one of promising software technologies
for achieving this goal. This note discusses rule-based approaches to automated
negotiation and presents some experimental results based on our own imple-
mentation of a rule-based price negotiation mechanism in a model e-commerce
multi-agent system. The experimental scenario considers multiple buyer agents
involved in multiple English auctions that are performed in parallel.

1 Introduction
During last few years, interest in e-commerce has shifted from simple Web presence of
businesses to advanced use of e-commerce technologies. These technologies are used
to support the growth of the business itself, by improving its efficiency and profitability.
Therefore, the idea of automating e-commerce transactions attracted a lot of research.
Currently, e-commerce systems are based on humans to make the most important
decisions in various activities along the lifeline of an e-commerce transaction. Software
agents are often claimed to be one of the best technologies for automating e-commerce
processes. In this context, we have set up a project to contribute to bridging the gap
between these two scenarios (real and possible). Our project has two main goals: i) to
build a large-scale implementation approximating an e-commerce environment; ii) to
develop a tool that can be used for modeling various e-commerce scenarios.
E-commerce utilizes (to various degrees) digital technologies to mediate commer-
cial transactions. As a part of our project we have conceptualized a commercial trans-
action as consisting of four phases: i) pre-contractual phase including activities like
need identification, product brokering, merchant brokering, and matchmaking; ii) ne-
gotiation where participants negotiate according to the rules of the market mechanism
and using their private negotiation strategies; iii) contract execution including activities
like: order submission, logistics, and payment; iv) post-contractual phase that includes
activities like collecting managerial information and product or service evaluation. Fo-
cus of this paper is on negotiation phase in a multi-agent e-commerce system that we
been developing and implementing ([7], [1]). As part of this work we are interested in
2

endowing our agents with flexibility necessary to engage in unknown in advance forms
of negotiations using rule-based approaches. Therefore we start with a brief overview
of background work on rule representations in automated negotiation. Then we outline
the design of our system that uses rules for enforcing specific negotiation mechanisms,
together with a sample experimental scenario. Finally, we discuss experimental results
obtained using our implementation (see [2] for implementation details).

2 Overview of Rule-Based Negotiation

Rules have been indicated as a promising technique for formalizing multi-agent negoti-
ations ([3, 4, 6, 11, 13, 16, 2]). Note that when designing systems for automated negotia-
tions one should distinguish between negotiation protocols (or mechanisms) that define
”rules of encounter” between participants and negotiation strategies that define behav-
iors aiming at achieving a desired outcome. Thus far multiple rule representations were
proposed for both negotiation mechanisms ([3, 14, 11, 16]) and strategies ([6, 13, 4]).
In our work we follow a rule-based framework for enforcing specific negotiation
mechanisms inspired by [3]. Its authors sketched a complete framework for imple-
menting portable agent negotiations that comprises: (1) negotiation infrastructure, (2)
generic negotiation protocol and (3) taxonomy of declarative rules. The negotiation
infrastructure defines roles of negotiation participants and of a host. Participants ex-
change proposals within a negotiation locale managed by the host. The generic nego-
tiation protocol defines three phases of a negotiation: (1) admission, (2) exchange of
proposals, and (3) formation of an agreement, in terms of how and when messages
should be exchanged between the host and participants. Negotiation rules are used for
enforcing the negotiation mechanism. Rules are organized into taxonomy: rules for par-
ticipants admission to negotiations, rules for checking validity of proposals, rules for
protocol enforcement, rules for updating negotiation status and informing participants,
rules for agreement formation and rules for controlling negotiation termination.
The proposal for formalizing negotiations introduced in [14] goes beyond the frame-
work of [3]. Its authors suggest use of an ontology for expressing negotiation protocols.
Whenever an agent is admitted to negotiation it also obtains a specification of the ne-
gotiation rules in terms of the shared ontology. In some sense, the negotiation template
used by our implementation (see [2]) is a ”simplified” negotiation ontology and the
participants must be able to ”understand” parameters defined in the template. This ap-
proach is exemplified with a sample scenario, but [14] contains neither implementation
details, nor experimental results.
In [16] a mathematical characterization of auction rules for parameterizing the auc-
tion design space is introduced. The proposed parametrization is organized along three
axes: i) bidding rules – state when bids may be posted, updated or withdrawn; ii) clear-
ing policy – states how the auction commands resource allocation (including auctioned
items and money) between auction participants (this corresponds roughly to agreement
making in our approach); iii) information revelation policy – states how and what inter-
mediate auction information is supplied to participating agents.
In [11] an implementation of a new rule-based language for expressing auction
mechanisms – AB3D scripting language — is reported. The design and implementa-
 3

tion of AB3D were primarily influenced by the parametrization of auction design space
defined in [16] and the previous experiences with the Michigan Internet AuctionBot
([15]). According to [11], A3BD allows the initialization of auction parameters, the
definition of rules for triggering auction events, declaration of user variables and defi-
nition of rules for controlling bid admissibility.
A formal executable approach for defining strategy of agents participating in nego-
tiations using defeasible logic programs is reported in [8] and [6]. This approach was
demonstrated using English auctions and bargaining with multiple parties by indicating
sets of rules for describing strategies of participating agents. However, no implementa-
tion results were reported. It is also interesting to note that despite the fact that paper [8]
claims to address both protocol and strategy representation in defeasible logic, finally
only a simplistic example of a protocol for two bidding rounds auction is described,
while more detailed examples are provided only for strategy representation.
In [13] a preliminary implementation of a system of agents that negotiate using
strategies expressed in defeasible logic is described. The implementation is demon-
strated with a bargaining scenario involving one buyer and one seller agent. The buyer
strategy is defined by a defeasible logic program. Note that the implementation reported
in [13] builds on the architecture of negotiating agents previously introduced in [6].
The CONSENSUS system that enables agents to engage in combined negotiations
was presented in [4]. CONSENSUS allows agents to negotiate different complemen-
tary items on separate servers on behalf of human users. Each CONSENSUS agent
uses rules partitioned into: i) basic rules that determine the negotiation protocol, ii)
strategy rules that determine the negotiation strategy and iii) coordination rules that de-
termine knowledge for assuring that either all of the complementary items or none are
purchased. Note that in CONSENSUS the rule-based approach is taken beyond mecha-
nism and strategy representation to capture also coordination knowledge.

3 System Description and Experiment

Conceptual Architecture. Our system acts as a distributed marketplace in which agents

perform functions typically observed in e-commerce ([7]). E-shops and e-buyers are
represented by shop and seller, and respectively client and buyer agents. In the exper-
iments we considered simplified version of this scenario that involves a single shop
agent S and n client agents Ci , 1 ≤ i ≤ n. The shop agent is selling m products
P = {1, 2, . . . , m}. We assume that each client agent Ci , 1 ≤ i ≤ n, is seeking a set
Pi ⊆ P of products. Shop agent S is using m seller agents S j , 1 ≤ j ≤ m and each seller
agent S j is responsible for selling a single product j. Each client agent Ci is using buyer
agents Bik to purchase products from the set Pi . Each buyer agent Bik is responsible
with negotiating and buying exactly one product k ∈ Pi , 1 ≤ i ≤ n. To attempt purchase
buyer agents Bik migrate to the shop agent S and engage in negotiations; a buyer agent
Bik , that was spawned by client agent Ci , will engage in negotiation with seller S k , to
purchase product k. This simple scenario is sufficient for the purpose of our paper, i.e. to
illustrate how a number of rule-based automated negotiations can be performed concur-
rently. In this setting, each seller agent S j plays the role of a negotiation host defined in
[3]. Therefore, in our system, we have exactly m instances of the framework described
4

in [3]. Each instance is managing a separate ”negotiation locale,” while all instances are
linked to the shop agent S . For each instance we have one separate set of rules together
with a negotiation template that describes the negotiation mechanism implemented by
that host. Note that each seller may use a different negotiation mechanism (different
form of an auction, or an auction characterized by different parameters, such as the
starting price or the bidding increment). See [2] for the details of our implementation
of this conceptual architecture using JADE ([9]) and JESS ([10]).
English Auctions and Participants Strategy. For the purpose of this paper we have
set our system up for a particular negotiation scenario involving English auctions. In an
English auction there is a single item sold by a single seller and many buyers bidding
against each other to buy this item. Usually, there is a time limit for ending the auction,
a seller reservation price that must be met by the winning bid for the item to be sold
and a minimum value of the bid increment. A new bid must be higher than the currently
highest bid plus a minimal bid increment in order to be accepted. All the bids are visible
to all auction participants.
The constraints describing English auctions were encoded as a modularized set of
JESS rules. The rules were then used to initialize the negotiation hosts [2]).
Strategies of participant agents are defined in accordance with the negotiation pro-
tocol (i.e. English auction in this particular setting). Basically, strategy defines if and
when a given participant will submit a proposal and what are its parameters. For the
time being we opted for a simple solution: the participant submits a first bid imme-
diately after it is granted admission and whenever it gets a notification that another
participant issued an accepted proposal. The value of the bid is equal with the sum of
the currently highest bid and an increment value that is private to the participant. Each
participant has its own valuation of the negotiated product. If the value of the new bid
exceeds this value then the proposal submission is canceled.
Strategies were implemented in Java as behaviors of participant JADE agent ([9]).
In the future we plan to investigate and experiment also with rule-based representations
of strategies ([6, 8]). This will provide us with the required flexibility and allow to easily
add multiple strategies to our implementation.
Experiment. In the experiment we considered m = 10 products and n = 12 clients
seeking all of them, i.e. Pi = P for all 1 ≤ i ≤ 10. Auction parameters were the same for
all auctions: reservation price 50 and the minimum bid increment 5. Clients reservation
prices were randomly selected from the interval [50,72] and their bid increments were
randomly selected from the interval [7,17].
In this experiment 143 agents were created: 1 shop S , 10 sellers S i , 1 ≤ i ≤ 10, 12
clients Ci , 1 ≤ i ≤ 12, and 120 buyers Bik , 1 ≤ i ≤ 12, 1 ≤ k ≤ 10; and 10 English
auctions were run concurrently. One separate JESS rule engine was also created for
each English auction (therefore a total of 10 JESS rule engines were run in parallel). A
total of more than 1400 messages was exchanged during negotiations. This means that
the average number of messages exchanged per negotiation was about 140. While the
total number of agents and messages is still small (for instance in comparison with these
reported in [5]) these results indicate that the proposed approach has good potential for
supporting large-scale experiments. Figure 1 shows messages exchanged between seller
S 1 and buyers Bi1 , 1 ≤ i ≤ 12 that were captured with the JADE sniffer agent ([9]).
 5

Fig. 1. Negotiation of a seller with 12 buyers in an English auction

Analyzing the results we have noticed that all auctions finished successfully. Let us
comment on the result of auction for product 9. This auction finished with winner B12,9
and winning price 66. Clients Ci (1 ≤ i ≤ 12) reservation prices (i.e. maximum prices
clients would agree to pay for product 9) were: 71, 70, 70, 70, 70, 70, 70, 70, 50, 50, 70,
70 and their bid increments were: 10, 15, 8, 7, 9, 17, 17, 13, 11, 6, 16, 12. This means
that when agent B12,9 got its bid of 66 accepted, neither Bi,9 agent (i , 12) was able to
bid more without exceeding its own reservation price. For example, B4,9 would have to
bid 66 + 7 = 73 > 70 which was not acceptable according to B4,9 Šs strategy.
We have also noticed that many bids were rejected because the rule stating that ”a
new bid must be higher than the currently highest bid plus a minimal bid increment”
was violated. This might look incorrect because all buyers bid increments were set to
values greater than 5 (the minimum value of the bid increment required by the auctions).
The explanation is the race condition that occurs between competing buyer agents that
bid against each other almost simultaneously. When they decide to bid, agents compute
their bid correctly. However, when the first of the competing bids that reaches the host
is validated, the currently highest bid is updated and this update might invalidate other
incoming bids that didn’t reach the host on time.

4 Conclusions and Future Work

In this paper we discussed rule-based approaches to automated negotiations in multi-
agent e-commerce systems. The discussion was supplemented by providing experimen-
tal results obtained using our own implementation of a rule-based price negotiation
framework. These results support the claim that rules are a feasible and scalable tech-
nology for approaching flexible automated negotiations in e-commerce.
6

As future work we plan to: (i) complete the integration of the rule-based framework
into our e-commerce model ([1]); (ii) asses the generality of our implementation by ex-
tending it to include other price negotiation mechanisms and multiple strategy modules.
We will report on our progress in subsequent papers.

References
1. Bădică, C., Ganzha, M., Paprzycki, M.: UML Models of Agents in a Multi-Agent E-
Commerce System. In: Proc. ICEBE’2005, Beijing, China. IEEE Computer Society Press,
Los Alamitos, CA, (2005) 56–61.
2. Bădică, C., Bădiţă, A., Ganzha, M., Iordache, A., Paprzycki, M.: Rule-Based Framework for
Automated Negotiation: Initial Implementation. In: A. Adi, S. Stoutenburg, S. Tabet (eds.):
Proc. RuleML’2005, Galway, Ireland. LNCS 3791, Springer Verlag (2005) 193–198.
3. Bartolini, C., Preist, C., Jennings, N.R.: A Software Framework for Automated Negotiation.
In: Proc. SELMAS’2004, LNCS 3390, Springer Verlag (2005) 213–235.
4. Benyoucef, M., Alj, H., Levy, K., Keller, R.K.: A Rule-Driven Approach for Defining the Be-
haviour of Negotiating Software Agents. In: J.Plaice et al. (eds.): Proceedings of DCW’2002,
LNCS 2468, Springer verlag (2002) 165–181.
5. Chmiel, K., Tomiak, D., Gawinecki, M., Karczmarek, P., Szymczak, Paprzycki, M.: Testing
the Efficiency of JADE Agent Platform. In: Proc. 3rd International Symposium on Parallel
and Distributed Computing, Cork, Ireland. IEEE Computer Society Press, Los Alamitos, CA,
USA, (2004), 49–57.
6. Dumas, M., Governatori, G., ter Hofstede, A.H.M., Oaks, P.: A Formal Approach to Negoti-
ating Agents Development. In: Electronic Commerce Research and Applications, Vol.1, Issue
2 Summer, Elsevier Science, (2002) 193–207.
7. Ganzha, M., Paprzycki, M., Pîrvănescu, A., Bădică, C., Abraham, A.: JADE-based Multi-
Agent E-commerce Environment: Initial Implementation, In: Analele Universităţii din Tim-
işoara, Seria Matematică-Informatică, Vol.XLII, (2004), 79–100.
8. Governatori, G., Dumas, M., ter Hofstede, A.H.M., and Oaks, P.: A formal approach to pro-
tocols and strategies for (legal) negotiation. In: Henry Prakken, editor, Proc. 8th International
Conference on Artificial Intelligence and Law, IAAIL, ACM Press, (2001) 168–177.
9. JADE: Java Agent Development Framework. See http://jade.cselt.it.
10. JESS: Java Expert System Shell. See http://herzberg.ca.sandia.gov/jess/.
11. Lochner, K.M., Wellman, M.P.: Rule-Based Specification of Auction Mechanisms. In: Proc.
AAMAS’04, ACM Press, New York, USA, (2004).
12. Lomuscio, A.R., Wooldridge, M., Jennings, N.R.: A classification scheme for negotiation in
electronic commerce. In: F. Dignum, C. Sierra (Eds.): Agent Mediated Electronic Commerce:
The European AgentLink Perspective, LNCS 1991, Springer Verlag (2002) 19–33.
13. Skylogiannis, T., Antoniou, G., Bassiliades, N.: A System for Automated Agent Negotiation
with Defeasible Logic-Based Strategies – Preliminary Report. In: Boley, H., Antoniou, G.
(eds): Proc. RuleML’04, Hiroshima, Japan. LNCS 3323 Springer-Verlag (2004) 205–213.
14. Tamma, V., Wooldridge, M., Dickinson, I: An Ontology Based Approach to Automated Ne-
gotiation. In: Proceedings AMEC’02: Agent Mediated Electronic Commerce, LNAI 2531,
Springer-Verlag (2002) 219–237.
15. Wurman, P.R., Wellman, M.P., Walsh, W.E.: The Michigan Internet AcutionBot: A Con-
figuarable Auction Server for Human and Software Agents. In: Proceedings of the second
international conference on Autonomous agents. Agents’98, Minneapolis, USA. ACM Press,
New York, USA, (1998), 301-308.
16. Wurman, P.R., Wellman, M.P., Walsh, W.E.: A Parameterization of the Auction Design
Space. In: Games and Economic Behavior, 35, Vol. 1/2 (2001), 271–303.